This is my ongoing implementation of the paper by Gupta et al.
MSCOCO, PPDB, Wikianswers are available in this repo. I put source / target files into one file with the follwing command.
paste train_source_1.txt train_target_1.txt > train.txt
paste train_source.txt train_target.txt > valid.txt
paste test_source.txt test_target.txt > test.txt
cat train.txt valid.txt > merged.txt && mv merged.txt train.txt && rm valid.txt